hi niklas,
i'm curious about your implementation:
- have you been doing some profiling of the scheduling overhead?
- which latency settings have you been using? it would be great to know
the worst-case response times of the locking synchronization ...
in general, the expressive power of dataflow languages in terms of parallelism is really amazing, however neither pd nor nova are general-purpose programming languages, but low-latency soft-realtime audio programming languages, which makes a usable implementation rather complex ...
cheers, tim
On Thu, 2007-05-31 at 02:19 +0200, Niklas Klügel wrote:
Tim Blechmann wrote:
On Wed, 2007-05-30 at 12:13 +0200, Niklas Klügel wrote:
I think it depends on the application.... for the most part, we
can't
get a generic speedup from using multiple cores (forgive me if
wrong)
that would apply to every single pd program..... but some types of computations such as large ffts can be performed faster when distributed to different cores, in which case, the code for the fft has to be parallelized a priori. Plus, the memory is tricky. You
can
have a memory access bottleneck, when using a shared memory resource between multiple processors. It's definitely a problem that is worth solving, but I'm not suggesting to do anything about it soon. It sounds like something that would require a complete top-down re-design to be successful. yikes
Chuck
I once wrote such a toolset that does automatically scale up with multiple threads throughout the whole network. it worked by detecting cycles in the graph and splits of the signals while segmenting the graph in autonomous sequential parts and essentially adding some smart and lightweight locks everyhwere the signals split or merged. it even reassigned threats on the lock-level to "balance" the workload in the graph and preventing deadlocks. the code is/was around 2.5k lines of c++ code and a bloody mess :) so, i don't know much about the internals of pd but it'd be probably possible.
detaching ffts (i.e. canvases with larger blocksizes than 64) should be rather trivial ...
distributing a synchronous dsp graph to several threads is not trivial, especially when it comes to a huge number of nodes. for small numbers of nodes the approach of jackdmp, using a dynamic dataflow scheduling, is probably usable, but when it comes to huge dsp graphs, the synchronization overhead is probably to big, so the graph would have to be split to parallel chunks which are then scheduled ...
true, i didn't try big graphs, so i can't really say how it would behave. it was more a fun project to see if it was doable. at that time i had the impression that the locking and the re-assignment of threads was quite efficient and done only on demand, if the graph has more sequential parts than the number of created threads ; i am curious how it can be achieved in a lock-free way.
about the issues of explicitely threading parts of the graph (that came up in the discussion lateron), i must say i don't get why you would want to do it. seeing how the numbers of cores are about to increase, i'd say that it is contraproductive in relation to the technological development of hardware and the software running on top of it lagging behind as well as the steady implicit maintenance of the software involved. from my point of view a graphical dataflow language has the perfect semantics to express the parallelisms of a program in an intuitive way. therefore i'd say that rather than adding constructs for explicit parallelism to the language that is able to express them anyhow adding constructs for explicit serialization of a process makes more sense. maybe i'm talking nonsense here, please correct me.
so long... Niklas
PD-list@iem.at mailing list UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list
-- tim@klingt.org ICQ: 96771783 http://tim.klingt.org
Every word is like an unnecessary stain on silence and nothingness Samuel Beckett